Titel: Generative Modeling via Flows on Measure Spaces
Abstract: Generative models with neural network approximations have shown impressive results in many applications, in particular in inverse problems and data assimilation. Among these models, flow matching techniques stand out for their simple applicability and good scaling properties. We give an introduction into flow matching and show some results of our group in this direction, namely
i) an extension to generator matching of Markov processes to learn time series,
ii) an incorporation of flow matching into plug-and-play algorithms to solve image restoration problems
iii) conditional flow matching using conditional Wasserstein distances, and
iv) telegrapher's generative modeling via Kac flow.
Zur Vortragenden: Gabriele Steidl ist Professorin für angewandte Mathematik an der Technischen Universität Berlin und eine der führenden Experten zu mathematischen Aspekten der Bildverarbeitung, wobei Ihre Arbeiten unter anderem signifikante Beiträge zur harmonischen Analysis mit Fokus auf effizienten Rechenmethoden, sowie konvexer Optimierung und maschinellem Lernen umschließen. Ihre Auszeichnungen und Aktivitäten in der mathematischen Gesellschaft beinhalten die Auswahl zum Fellow der Society for Industrial and Applied Mathematics, das Arbeiten als Editor-in-Chief des SIAM Journal on Imaging Sciences, sowie Mitgliedschaften im Executive Boards des MATH+ Exzellenzclusters in Berlin sowie des Scientific Advisory Committees der Helmholtz Imaging Platform der Helmholtz-Gemeinschaft.